Multimedia Answer Generation for Community Question Answering

1. Multimedia Answer Generation for Community Question Answering

2. Problem Statement • Textual Answers • Multimedia Answers

3. Literature Survey

4. System Decomposition • Answer medium selection, • Query generation • Multimedia data selection and presentation.

5. Pre-requisites • Datasets • Image & Video Mining API • Flickr, Picasaweb, Youtube, etc.

6. Process Flow Dataset collection Multimedia Data Duplicate Removal Classification(Conversational & Informational) Result Re-ranking The answer medium selection and query selection components Query Generation for Multimedia Retrieval

7. Answer Medium Selection • Classification • only text, • Text + image • text + video • text + image + video • Approach: • Question Based Classification • Answer Based Classification • Media Resource Analysis

8. Query Extraction For each QA pair, we generate three queries. 1. Convert the question to a query, 2. Identify several key concepts from verbose answer which will have the major impact on effectiveness. 3. Finally, we combinethe two queries that are generated from the question and the answer respectively.

9. Query Generation for Multimedia Search • Query Extraction • The second step is query selection.

10. Query Selection • Three-class classification task, since we need to choose one from the three queries • We adopt the following features: • POS Histogram. • For the queries that contain a lot of complex verbs it will be difficult to retrieve meaningful multimedia results. • We use POS tagger to assign part-of-speech to each word of both question and answer. • Here we employ the Stanford Log-linear Part-Of-Speech Tagger and 36 POS are identified. • We then generate a 36-dimensional histogram, in which each bin counts the number of words belonging to the corresponding category of part-of-speech.

11. (2) Search performance prediction. • Clarity score for each query based on the KL divergencebetween the query and collection language models. • Wecan generate 6-dimensional search performance prediction featuresin all (there are three queries and search is performed on both image and video search engines). • Therefore, for each QA pair, we can generate 42-dimensional features. • Based on the extracted features, we train an SVM classifier with a labeled training set for classification • i.e., selecting one from the three queries.

12. Clarity function:

13. Multimedia Data Selection & Prediction • We perform search using the generated queries to collect image and video data with Google image and video search engines respectively. • Most of the current commercial search engines are built upon text-based indexing and usually return a lot of irrelevant results. • Therefore, • Re-ranking by exploring visual information is essential to reorder the initial text-based search results. • Here we adopt the graph-based re-ranking method.

14. Graph Based Re-Ranking & Duplicate Removal

15. List of Algorithms • Core sentence extraction from question • Stemming & stop-words removal on answers • Question Type based on Answer Medium ( Naïve Bayes) • Head word extraction • Media Resource Analysis • Clarity score based on KL Divergence • Query Generation • Query Selection • POS Feature Extraction • Search Performance Prediction • Multimedia Data selection & presentation • Graph based ranking • Face Detection Algorithms • Feature Extraction from images • Key frame identification & extraction

16. References • M. Surdeanu, M. Ciaramita, and H. Zaragoza, “Learning to rank answers on large online QA collections,” in Proc. Association for Computational Linguistics, 2008 • S. Cronen-Townsend, Y. Zhou, andW. B. Croft, “Predicting query performance,” in Proc. ACM Int. SIGIR Conf., 2002. • LiqiangNie, MengWang, YueGao, Zheng-Jun Zha, and Tat-SengChua“Beyond Text QA: Multimedia Answer Generation by Harvesting Web Information” IEEE Multimedia Transaction 2013

17. Thank You….